Artificial intelligence is rapidly becoming part of biotechnology. Algorithms can now predict protein structures, identify possible therapeutic targets, screen millions of virtual compounds and generate entirely new molecular designs.
These advances have created hopes that AI could make drug development faster, less expensive and more successful. In 2025, researchers reported results from a phase 2a clinical trial of a small-molecule drug whose target and chemical structure were identified with the assistance of generative AI. AI-designed proteins have also been built and tested in laboratories, demonstrating that algorithms can create functional molecules that do not exist in nature.
However, predicting a promising molecule is not the same as developing a medicine. An AI model cannot determine from computation alone whether a drug will be safe, reach the correct tissue, remain stable in the body or improve patient outcomes.
The most realistic role for AI is therefore not to replace biological research. It is to help scientists decide which experiments are most worth performing.
Why drug discovery is difficult
Traditional drug development begins by identifying a biological target involved in a disease. Researchers then search for a molecule capable of modifying that target in a useful way.
A candidate must satisfy many requirements. It may need to:
- Bind the intended target
- Produce the desired biological effect
- Avoid important unintended targets
- Reach the relevant tissue
- Remain stable long enough to work
- Be absorbed or delivered effectively
- Avoid toxic metabolites
- Be manufactured consistently
- Provide benefits that outweigh its risks
A compound can perform well in one category and fail in another. A highly potent molecule may be too toxic. A drug that works in isolated cells may be broken down rapidly by the liver. A treatment that performs well in animals may not work in humans.
Thousands of compounds may be considered during early discovery, but only a small percentage enter clinical trials—and only a fraction of those are eventually approved. AI aims to improve this process by identifying patterns in biological and chemical data that would be difficult for researchers to detect manually.
Protein structure is central to biology
Proteins are chains of amino acids that fold into three-dimensional structures. Their shapes influence how they function and interact with other molecules.
A protein’s structure can reveal:
- Where a drug might bind
- How a mutation changes function
- How an enzyme recognizes its substrate
- How an antibody interacts with an antigen
- Which regions might be targeted therapeutically
Experimental methods such as X-ray crystallography, nuclear magnetic resonance spectroscopy and cryo-electron microscopy can determine protein structures. These techniques are essential but may require significant time, specialized equipment and carefully prepared samples.
Predicting structure from an amino-acid sequence was a longstanding challenge because a protein can theoretically adopt an enormous number of possible conformations.
AlphaFold and structure prediction
AI changed protein-structure prediction dramatically. AlphaFold 2 demonstrated that deep-learning systems could predict many protein structures with accuracy approaching that of experimental methods for suitable targets.
AlphaFold 3 expanded the scope beyond isolated proteins. The model can predict joint structures involving proteins, nucleic acids, small molecules, ions and modified residues. In published benchmarking, it improved the prediction of several protein–ligand, protein–nucleic-acid and antibody–antigen interactions compared with many previous specialized methods.
This can help researchers:
- Generate hypotheses about protein function
- Interpret genetic variants
- Identify possible binding pockets
- Model protein complexes
- Plan laboratory experiments
- Prioritize targets for drug development
AlphaFold does not “solve biology,” however. Its output is a prediction based on patterns learned from existing data.
A predicted structure is not an experimental structure
Proteins are not rigid objects. They move between multiple conformations, respond to their environment and interact with other molecules. A static structure may capture one likely arrangement while missing biologically important motion.
Prediction can also be difficult when:
- A protein is intrinsically disordered
- Multiple conformations are functionally important
- The protein interacts with a membrane
- Binding requires a major structural change
- Unusual chemical modifications are present
- The target has few related examples in the training data
Confidence scores can help researchers assess the reliability of a prediction, but they do not prove that every atomic position is correct. A high-confidence protein structure also does not guarantee an accurate prediction of binding strength, reaction rate or therapeutic response.
Experimental structural biology remains necessary, especially when decisions depend on precise atomic interactions.
From prediction to generative biology
Structure-prediction models answer a question such as: “What shape is this protein sequence likely to adopt?”
Generative models attempt to answer a different question: “What new sequence or structure could produce the function we want?”
Generative biology applies machine learning to create new:
- Protein sequences
- Protein structures
- Antibodies
- Enzymes
- Peptides
- Regulatory DNA or RNA sequences
- Small drug-like molecules
Some models function similarly to language models. Instead of learning relationships among words, they learn patterns among amino acids, molecular fragments or biological sequences.
Other systems use diffusion models. These algorithms begin with a noisy or random representation and gradually transform it into a structured molecular design that meets selected constraints.
Designing proteins that do not exist in nature
RFdiffusion is one example of a generative protein-design system. It was trained to generate three-dimensional protein backbones, including structures designed to bind specific molecular targets.
Researchers used RFdiffusion to create protein structures with selected shapes, symmetries and binding surfaces. They then converted those structures into amino acid sequences, manufactured the proteins and tested them experimentally.
Importantly, the laboratory testing was not an optional final detail. Many computer-generated designs will fail to fold, remain stable or bind their intended targets. Experimental screening identifies the minority that behaves as predicted.
Generative protein design may eventually support the development of:
- Therapeutic proteins
- Antibodies
- Vaccine antigens
- Enzyme replacements
- Molecular sensors
- Diagnostic reagents
- Targeted drug-delivery systems
It may also improve industrial enzymes used in food production, manufacturing, environmental remediation and sustainable chemistry.
How AI may improve antibody development
Antibodies bind molecular targets through complex three-dimensional surfaces. Traditional antibody discovery frequently involves immunizing animals, screening large molecular libraries, or modifying an existing antibody through repeated laboratory cycles.
AI models can propose antibody sequences predicted to bind a specific region of an antigen. They may also help optimize:
- Binding affinity
- Specificity
- Stability
- Solubility
- Manufacturing yield
- Risk of aggregation
- Potential immune recognition
This could reduce the number of candidates that must be produced and tested. However, antibody behavior is influenced by factors that remain difficult to predict, including conformational changes, glycosylation, immune-system interactions and performance in living organisms.
A computationally impressive antibody still requires biochemical testing, cell studies, animal models when appropriate, and clinical trials.
Designing small-molecule drugs
Most traditional medicines are small molecules capable of interacting with proteins or other biological targets.
AI can support several stages of small-molecule discovery:
- Identifying disease-associated targets
- Predicting possible binding sites
- Screening virtual chemical libraries
- Generating new molecular structures
- Estimating solubility and permeability
- Predicting metabolism
- Flagging possible toxicity
- Optimizing several properties simultaneously
Conventional virtual screening may evaluate a collection of known or proposed molecules against a target. Generative chemistry goes further by creating new structures predicted to satisfy selected requirements.
For example, a model might be asked to generate compounds that bind a protein pocket, remain within a desired molecular-weight range, avoid a known toxicity pattern, and possess properties compatible with oral administration.
The result is not a finished drug. It is a prioritized set of hypotheses.
Candidate screening: narrowing the search
One of AI’s clearest benefits is reducing the number of candidates that require physical testing.
A traditional screen might test hundreds of thousands of compounds in a laboratory. A computational model could first rank a virtual library containing millions or billions of possible structures. Scientists could then synthesize and test a smaller group of highly ranked candidates.
AI-based screening can incorporate several types of information:
- Molecular structure
- Gene-expression data
- Protein interactions
- Cell-imaging data
- Genetic screens
- Electronic health records
- Published scientific literature
- Known drug responses
Models can also support drug repurposing by identifying approved compounds that may influence a newly identified disease pathway. Because these compounds already have some human safety information, repurposing may shorten parts of development—but clinical evidence is still required for the new use.
Can AI discover new therapeutic targets?
AI is also being used before molecule design begins.
Disease biology often involves networks of genes, proteins, immune signals and cell types. Machine-learning models can analyze large multi-omics datasets to identify relationships among genetic variants, gene expression, proteins, metabolites and clinical outcomes.
A model may suggest that a previously overlooked protein is strongly associated with disease progression. Researchers can then test whether changing that protein’s activity actually affects the disease.
This distinction is crucial: association does not establish causation. A protein may change because disease is present without being a useful therapeutic target. Laboratory experiments are needed to determine whether manipulating it produces a beneficial effect.
The first phase 2 trial of a generative-AI drug
Rentosertib, previously known as ISM001-055, is an investigational small-molecule inhibitor of TNIK being studied for idiopathic pulmonary fibrosis.
The target and molecular candidate were identified using AI-assisted platforms. In 2025, researchers published results from a randomized phase 2a trial involving 71 participants who received one of three rentosertib regimens or placebo for 12 weeks.
The study primarily evaluated safety and tolerability. Adverse events occurred across both treatment and placebo groups, and the trial reported exploratory signals related to lung function. However, it was small and short, and it was not designed to establish definitive clinical effectiveness.
The study was nevertheless significant because it demonstrated that an AI-assisted drug-design program could progress beyond target identification and early testing into a randomized phase 2 clinical trial.
It did not demonstrate that AI-developed drugs are generally more effective than conventionally discovered medicines. That question will require evidence from many candidates, comparison groups and complete development programs.
What AI predictions can miss
Biological systems are context-dependent. A molecular interaction predicted in a computer may behave differently inside a cell, tissue or patient.
Important limitations include:
- Training-data bias
- AI models learn from existing datasets. These data may overrepresent well-studied protein families, chemical structures, diseases or populations. Poorly studied biology may receive less reliable predictions.
- Incorrect or incomplete data
- Experimental databases contain measurement errors, inconsistent methods and negative results that may never have been published. Models can reproduce these limitations.
- Correlation without mechanism
- An algorithm may identify a strong statistical relationship without explaining whether it is biologically causal.
- Static representations
- Many models treat proteins as relatively stable structures, even though biological molecules move and change shape.
- Uncertain toxicity
- Toxicity may result from complex interactions among organs, metabolites, immune responses and genetic differences. These effects remain difficult to predict reliably from molecular structure alone.
- Lack of real-world context
- A candidate may work in a purified biochemical assay but fail in a cell because it cannot cross the membrane. It may work in cells but fail in an organism because it is cleared too rapidly or accumulates in the wrong tissue.
Synthesizability matters
Generative models can propose molecules that appear attractive computationally but are difficult or impossible to manufacture.
A proposed compound may require:
- Unstable intermediates
- Impractical reaction conditions
- Too many synthesis steps
- Rare starting materials
- Difficult purification
- Poor manufacturing reproducibility
Modern models increasingly include synthetic feasibility in their optimization criteria. Automated chemistry platforms can also attempt to plan reaction pathways and carry out repeated design–make–test cycles.
Even so, a virtual molecule has no therapeutic value until it can be produced, purified and characterized.
Laboratory validation is the essential checkpoint
AI can suggest what might work. Experiments determine whether it does.
A typical validation pathway may include:
- Synthesis or biological production of the candidate
- Confirmation of molecular identity and purity
- Measurement of target binding
- Testing in biochemical assays
- Evaluation in relevant cell systems
- Assessment of selectivity against other targets
- Pharmacokinetic and toxicity studies
- Animal testing when scientifically and ethically appropriate
- Human clinical trials
- Regulatory review and post-approval monitoring
Each stage can expose a weakness that was not visible computationally.
AI may also improve the experiments themselves. Models can recommend which compounds to test next, analyze high-content microscopy images or interpret multi-omics responses. Results can then be fed back into the model, creating an iterative cycle.
This closed-loop system—design, build, test and learn—may be more transformative than any single algorithm.
Laboratory automation and self-driving laboratories
Some biotechnology laboratories are integrating AI with robotics and automated equipment.
A self-driving laboratory can potentially:
- Select an experiment
- Prepare samples
- Run the assay
- Analyze the results
- Update the model
- Design the next experiment
This may allow researchers to explore molecular possibilities faster and with greater consistency. Automated systems are particularly useful for repetitive processes such as protein expression, compound screening and optimization of experimental conditions.
Human oversight remains necessary. Scientists must define meaningful objectives, identify flawed measurements, interpret unexpected findings and determine whether the model is optimizing something biologically useful.
An automated system can efficiently pursue the wrong goal if its objective is poorly designed.
Safety, transparency and reproducibility
AI-assisted drug development raises questions about how models should be evaluated and regulated.
Developers may not disclose complete training datasets, model architecture or optimization criteria because of intellectual-property concerns. This can make independent replication difficult.
Key questions include the following:
- What data were used to train the model?
- Were test data kept separate from training data?
- How does performance compare with existing methods?
- Are uncertainty estimates calibrated?
- Can independent researchers reproduce the results?
- Were unsuccessful predictions reported?
- How are human biases represented in the data?
- Who is responsible when a model-guided decision fails?
Regulators do not approve an algorithm’s confidence score as evidence that a drug works. The resulting therapy must meet the same standards for manufacturing quality, safety and clinical effectiveness as any other medicine.
AI will not remove clinical failure
Much of drug-development failure occurs because human biology is more complicated than laboratory models suggest. A target may be biologically valid but clinically unimportant. A treatment may affect the pathway without changing symptoms, disability or survival.
AI may reduce some early failures by improving target selection and molecular design. It may also help identify toxicity earlier. However, it cannot fully predict how diverse patients will respond over months or years.
Clinical trials remain necessary because they answer questions that cannot be resolved by molecular modeling:
- Does the treatment improve meaningful health outcomes?
- Which patients benefit?
- How large is the benefit?
- What adverse effects occur?
- How does it compare with existing care?
- Does the effect remain durable?
A faster route to a clinical candidate is valuable, but it does not eliminate the need for careful clinical development.
The bottom line
AI is changing biotechnology by improving protein-structure prediction, generating new proteins and chemicals, prioritizing therapeutic targets and narrowing enormous candidate libraries.
Tools such as AlphaFold 3 and RFdiffusion demonstrate that algorithms can model complex biological structures and create proteins with designed features. The progress of rentosertib into a randomized phase 2a trial shows that AI-assisted drug discovery can produce candidates suitable for human testing.
The strongest evidence, however, does not support the idea that algorithms can independently invent finished medicines. AI produces predictions and candidates—not proof.
The future of AI-enabled biotechnology will depend on combining computational speed with experimental discipline. Algorithms can explore molecular possibilities at a scale no human team could match, while laboratory studies and clinical trials determine which possibilities are biologically real, medically useful and acceptably safe.
This article is intended for scientific education and does not replace medical advice. AI-generated therapeutic candidates require laboratory, preclinical, clinical and regulatory validation before they can be considered safe or effective treatments.
References
Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ronneberger, O., Willmore, L., Ballard, A. J., Bambrick, J., Bodenstein, S. W., Evans, D. A., Hung, C.-C., O’Neill, M., Reiman, D., Tunyasuvunakool, K., Wu, Z., Žemgulytė, A., Arvaniti, E., . . . Jumper, J. M. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493–500. https://doi.org/10.1038/s41586-024-07487-w
Anishchenko, I., Pellock, S. J., Chidyausiku, T. M., Ramelot, T. A., Ovchinnikov, S., Hao, J., Bafna, K., Norn, C., Kang, A., Bera, A. K., DiMaio, F., Carter, L., Chow, C. M., Montelione, G. T., & Baker, D. (2021). De novo protein design by deep network hallucination. Nature, 600, 547–552. https://doi.org/10.1038/s41586-021-04184-w
Cao, L., Coventry, B., Goreshnik, I., Huang, B., Sheffler, W., Park, J. S., Jude, K. M., Marković, I., Kadam, R. U., Verschueren, K. H. G., Verstraete, K., Walsh, S. T. R., Bennett, N., Phal, A., Yang, A., Kozodoy, L., DeWitt, M., Picton, L., Miller, L., . . . Baker, D. (2022). Design of protein-binding proteins from the target structure alone. Nature, 605, 551–560. https://doi.org/10.1038/s41586-022-04654-9
Vázquez Torres, S., Leung, P. J. Y., Venkatesh, P., Lutz, I. D., Hink, F., Huynh, H.-H., Becker, J., Yeh, A. H.-W., Juergens, D., Bennett, N. R., Hoofnagle, A. N., Huang, E., MacCoss, M. J., Expòsit, M., Lee, G. R., Bera, A. K., Kang, A., De La Cruz, J., Levine, P. M., . . . Baker, D. (2024). De novo design of high-affinity binders of bioactive helical peptides. Nature, 626, 435–442. https://doi.org/10.1038/s41586-023-06953-1
Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J., Eisenach, H. E., Ahern, W., Borst, A. J., Ragotte, R. J., Milles, L. F., Wicky, B. I. M., Hanikel, N., Pellock, S. J., Courbet, A., Sheffler, W., Wang, J., Venkatesh, P., Sappington, I., Vázquez Torres, S., . . . Baker, D. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620, 1089–1100. https://doi.org/10.1038/s41586-023-06415-8
Xu, Z., Ren, F., Wang, P., Cao, J., Tan, C., Ma, D., Zhao, L., Dai, J., Ding, Y., Fang, H., Li, H., Liu, H., Luo, F., Meng, Y., Pan, P., Xiang, P., Xiao, Z., Rao, S., Satler, C., . . . Zhavoronkov, A. (2025). A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: A randomized phase 2a trial. Nature Medicine, 31, 2602–2610. https://doi.org/10.1038/s41591-025-03743-2
